10 papers
TailSieve: Partial-Rollout-Guided Tail Routing for LLM Rollouts
Tianqi Xu, Lu Lv, Haoyang Huang +15
Large-scale rollouts have become a core component of modern LLM systems, spanning reinforcement learning (RL) post-training, on-policy distillation (OPD), and sampling-heavy evalua…
OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs
Haoyang Huang, Wenjie Huang, Tianqi Xu +14
Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory…
Draft Less, Retrieve More: Hybrid Tree Construction for Speculative Decoding
Yuhao Shen, Tianyu Liu, Xinyi Hu +9
Speculative decoding (SD) accelerates large language model inference by leveraging a draft-then-verify paradigm. To maximize the acceptance rate, recent methods construct expansive…
When Hidden States Drift: Can KV Caches Rescue Long-Range Speculative Decoding?
Tianyu Liu, Yuhao Shen, Xinyi Hu +8
Speculative decoding accelerates LLM inference, but SOTA hidden-state-based drafters suffer from long-range decay: draft accuracy degrades as the speculative step increases. Existi…
Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects
Jun Zhang, Yicheng Ji, Feiyang Ren +7
Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token d…
On the Step Length Confounding in LLM Reasoning Data Selection
Bing Wang, Rui Miao, Chen Shen +7
Large reasoning models have recently demonstrated strong performance on complex tasks that require long chain-of-thought reasoning, through supervised fine-tuning on large-scale an…
See the Forest for the Trees: Loosely Speculative Decoding via Visual-Semantic Guidance for Efficient Inference of Video LLMs
Yicheng Ji, Jun Zhang, Jinpeng Chen +4
Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency during autoregressive generation. Speculative Decoding (SD) mitigates t…
HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference
Bowen Zeng, Feiyang Ren, Jun Zhang +4
Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches…
KNN-SSD: Enabling Dynamic Self-Speculative Decoding via Nearest Neighbor Layer Set Optimization
Mingbo Song, Heming Xia, Jun Zhang +4
Speculative Decoding (SD) has emerged as a widely used paradigm to accelerate the inference of large language models (LLMs) without compromising generation quality. It works by eff…
SWIFT: On-the-Fly Self-Speculative Decoding for LLM Inference Acceleration
Heming Xia, Yongqi Li, Jun Zhang +2
Speculative decoding (SD) has emerged as a widely used paradigm to accelerate LLM inference without compromising quality. It works by first employing a compact model to draft multi…